Mindroot

AI agents start every session from zero: architecture, conventions, commands, and hard-won gotchas get re-discovered from scratch, burning tokens, time, and money on knowledge that already existed yesterday — and the same recurring procedures get re-explained on every task.

Existing fixes fail both ways: vector databases are opaque black boxes no human can edit or audit, while scattered Markdown files have no retrieval at all. Teams need know-how that is searchable, versioned, and reusable — and a memory layer that stays strictly project-scoped, local-first, and open to inspection.

Mindroot fuses two layers into one memory: retrieval-optimized memories ranked by hybrid cosine + BM25 search point into human-editable Markdown notes that remain the durable source of truth — plain files you own and can read in any editor. Reads are self-healing: content-hash verification silently re-indexes files edited outside the tool, and `update_section` performs surgical edits that leave the rest of the file byte-identical. On top sits a skills layer that stores proven procedures under stable names with their own search index, conflict-safe saves, and full revision history.

One tiny local service delivers it everywhere: 16 MCP tools for agents, a REST API for scripts, a CLI, a daemon you can start, stop, and health-check, and a dashboard with an interactive knowledge graph — wired into OpenCode, Claude Code, or Codex with a single `mindroot install` command. Localhost-only by default, with constant-time API-key auth on every request. Fully open source on GitHub — free to run, inspect, and self-host.